E146: The 92% AI Failure: Unmasking Enterprise's Trillion-Dollar Mistake

Episode Summary

Executive Summary: Invisible’s CEO argues the biggest AI gains will come not from replacing humans with off-the-shelf tools, but from redesigning workflows around AI and humans together. He says enterprise adoption is stalled by messy data, unclear quality standards, and weak production/evaluation frameworks. Invisible aims to solve this through training/evals, expert human feedback, and an AI-native process platform for enterprise modernization.

Main Topics: AI-native process redesign vs. human replacement: The CEO frames AI value as coming from clean-sheet redesign of processes with both humans and machines, rather than simply layering AI onto existing workflows or replacing workers outright. Invisible’s business and market position: Invisible is presented as a profitable, fast-growing AI company focused on training, evaluation, and enterprise workflow modernization, with a large expert labor marketplace behind it. Why enterprise AI adoption stalls: A major theme is that enterprise AI struggles to reach production because of poor data quality, lack of standardized evaluation, and uncertainty about what counts as 'good' output at scale. Training data, evals, and subject matter expertise: The company’s core offering is building custom evaluations and human feedback loops using highly screened experts for difficult domains where synthetic data is insufficient. DeepSeek and the economics of model building: The CEO downplays DeepSeek as a paradigm shift, arguing its costs and techniques mostly fit within the existing model-building playbook rather than fundamentally changing the economics. Invisible’s future as an AI-native application platform: Beyond training/evals, Invisible wants to become a modern ServiceNow-like platform for verticalized, configurable enterprise applications built around AI workflows. The limits of copilots and the need for process redesign: He argues that copilots mainly improve marginal productivity, but real gains require re-engineering end-to-end workflows and removing entire steps from processes.

Key Arguments: AI’s biggest value comes when companies redesign processes from scratch with AI in mind, rather than bolt AI onto legacy workflows. Humans remain essential because many enterprise tasks require domain expertise, judgment, and evaluation that models cannot reliably provide alone. Enterprise AI has not yet found a stable production paradigm; only a small share of models reach deployment. Data quality is a primary bottleneck: if data is messy or unstructured, AI efforts fail before model quality even matters. The hardest enterprise problem is not generating outputs but defining, scoring, and validating what 'good' looks like across thousands of outputs. Custom evals and expert human feedback are necessary for high-stakes use cases such as finance, law, and claims processing. DeepSeek is interesting but not revolutionary; its reported costs still fit the current model-building paradigm. Copilots help on the margins, but real efficiency gains require workflow redesign and AI-native process architecture. Invisible’s strategy is to deepen its training/evals business while evolving into a broader AI-native application development and services platform. The enterprise software market remains fragmented and slow to modernize, creating a large long-term opportunity for AI-driven infrastructure and applications.

Data Points: 2024 revenue: $134 million - Invisible ended 2024 at this level of revenue. Profitability: Profitable - The company reported profitable operations in 2024. Growth ranking: 3rd fastest growing AI business in America over the last three years - Used to describe Invisible’s growth trajectory. DeepSeek compute-cost estimate (viral story): $5 billion - Initial narrative cited for building the models. DeepSeek compute-cost estimate (later estimate): $1.6 billion - Later estimate referenced from FT and other sources. DeepSeek GPU count: 50,000 GPUs - Approximate compute scale mentioned for the model build. Enterprise model production rate: About 8% - Rough industry stat cited for models making it to production. Invisible talent acceptance rate: 1% acceptance rate - Refers to its expert pool for training and evaluation work. Expert pool composition: About 30% PhDs and masters - Describes the credential mix of the talent marketplace. Legacy software prevalence: 70% of software in America is over 20 years old - Used to support the case for long-duration enterprise modernization. Legacy modernization market concentration: No player more than 3% - Indicates how fragmented the application modernization market is. McKinsey Gen AI builds: 200–240 Gen AI builds - The CEO cites the volume of work handled while leading QuantumBlack Labs.

Pivotal Quotes: "“The material benefit you're going to see is when you clean sheet any process to be like, how would I design this process knowing all the AI tools I have from scratch?”" — Matt: Explaining the central philosophy behind AI-native transformation. "“When good AI meets bad data, the data usually wins.”" — Matt: Describing why enterprise AI initiatives often fail before deployment. "“AI has not really figured out what I'd call production paradigm yet.”" — Matt: Summarizing the core bottleneck in enterprise adoption and scaling.

Implications: Enterprise AI winners will likely be those who combine model access, data cleanup, human expertise, and workflow redesign. The next phase is less about flashy copilots and more about production-grade systems that work reliably in specific business contexts.